MiniMax H3 testing in Pencil

# Live

Launched October 22

October 22, 2026

Embee Joy Medrano

Anyone who has tried to change one thing in an AI video knows the frustration. The shot is almost right, but the product needs to be different, the scene needs to be at night, or the actor needs to be someone else. Regenerating from scratch means losing the camera move you liked. Editing in place usually means something else breaks.

We tested MiniMax H3 on exactly those jobs: seven scenarios taken from real paid media and production briefs. Some results were strong, and some showed clear limits. Here’s the full breakdown, so you know where it fits today.

The short version

MiniMax H3 is strongest when it can lean on a source video, and weakest when the change has to be global or the format has to go beyond 16:9 and 9:16. Five of the seven use cases are usable today with some care, and two aren’t a strong fit yet.

Use case	What worked	Watch out for	Use it today?
Multi-aspect-ratio adaptation	Near-identical 16:9 and 9:16 output; clean resize of a connected video	Only 16:9 and 9:16; camera drift; human realism	Not as a full solution
Product variant change	Animation, camera movement, realism	Swaps between similar-looking variants	Conditionally
Time of day change	Camera movement preserved	Lighting only, no real atmosphere shift	Not today
Character change	Performance, emotion and body language transfer	Camera and wardrobe drift between attempts	Conditionally
Animation change	Animation swap, camera movement	Reference vs start-frame ambiguity	Conditionally
Additional assets	Realistic additions without a reference image	Darker color, proportion and scale	Works, with caveats
Style and motion transfer	Style and motion carried onto the character	Proportions must be stated in the prompt	Yes, with explicit proportions

Strong fit today

These three tests followed the source video’s motion and performance closely.

Style and motion transfer MiniMax H3 carried a visual style and motion onto a character successfully, and the standard Pencil prompt format works well. The one condition is that proportion and scale aren’t handled automatically, so state the proportions you want in the prompt.

Recommend today? Yes, as long as the prompt spells out proportion and scale.

Change of character Replacing the actor while keeping the original performance is one of the most useful edits in production, and H3 did it, following the source’s movement, emotion, acting and body language closely. Consistency is the weak spot. Across two attempts with the same prompt, the camera movement shifted both times, and the clothing shifted in the second.

Recommend today? Conditionally. Character and performance transfer works well, but expect to iterate to lock the camera and wardrobe.

Change of animation H3 replaced animation elements inside a video as requested and kept the camera movement well. The issue is in the setup. The MiniMax video-gen node has both a “reference” and a “start frame” input, and the model seems unsure which one is the true subject. In some runs the output started entirely from the attached reference image instead of the source video (the product in the source video, in our test), even though the prompt named the source video.

Recommend today? Conditionally. Animation change and camera consistency work well, but the subject-source ambiguity needs careful handling in setup.

Good fit with some setup

Change of product variant Swapping a product’s variant inside an existing video (a different version, pack or color option) is a common request in FMCG and retail production. H3 captured the animation, camera movement and realism of the source well. The swap fails or doesn’t register when the source and target variants are visually close, because H3 struggles to re-render color when the two palettes are similar.

Recommend today? Conditionally. It works well for variants that look clearly different, but isn’t yet reliable when the source and target are close in color.

Additional assets Adding new elements into an existing video worked, and the additions looked realistic even without a reference image. The cost was side effects on the source: the overall video came out darker than the original, and inserted elements had proportion and scale problems relative to the scene.

Recommend today? Works, with caveats. Check the output against your source for brightness shifts and for the scale of anything you add.

Limited fit today

Multi-aspect-ratio adaptation for paid media The brief was to generate on-brand lifestyle footage and adapt it across the aspect ratios paid media placements need. This is the use case that matters most for ad production, and it’s where H3 hits its hardest limit: it only generates 16:9 and 9:16. Within those two, the results were good. H3 produced near-identical generations in both formats, and it resized a connected source video well, though the output can show slight compression artifacts at first. Camera movement sometimes deviated from the prompt, and human subjects can look noticeably AI-generated rather than photorealistic.

Recommend today? Not as a full solution. The format restriction is a hard blocker for the core promise of multi-aspect-ratio adaptation. One route worth trying: generate strong 16:9 and 9:16 base assets with H3, then use a Pencil resize or reformat step (crop-and-recompose or generative outpainting) to cover the other ad sizes.

Change of time of day Shifting a video from day to night or golden hour without a reshoot would save a lot of production time. H3 kept the same camera movement as the source, but the transformation was superficial. Lighting shifted slightly, while the sky, environment mood and color temperature stayed largely the same, even when the prompt named the target time of day.

Recommend today? Not today. The core transformation isn’t convincingly achieved.

Prompting tips

State proportions and scale explicitly. Style and motion transfer needs it, and inserted assets showed scale problems too.

When resizing a connected video, tell H3 not to compress, warp or force-resize it.

Be explicit and repeat camera movement instructions. It may reduce drift, but it didn’t fully fix it in our testing.

Check the reference and start-frame inputs. If the output starts from the attached image instead of your source video, the model has picked the wrong subject.

For character swaps, plan to iterate to lock the camera and wardrobe.

We’re continuing to run tests on MiniMax H3. If there are specific edits or use cases you’ve been struggling to crack, let us know in the comments, we want to know what’s on your list.

Comments (0)

Popular